技术规格

GLM-4.5-Air

GLM-4.5-Air is a high-efficiency large language model developed by Z.ai as part of the GLM-4.5 series. It is designed to bridge the gap between massive-scale foundation models and the practical constraints of on-device or mid-range cloud deployments. Optimized primarily for agent-oriented workflows, the model prioritizes reasoning, complex instruction following, and code generation. It functions as a versatile engine for autonomous agents capable of multi-step planning and tool invocation, making it a viable selection for developers building sophisticated digital assistants and automated software engineering pipelines.

Architecturally, the model utilizes a sparse Mixture-of-Experts (MoE) framework, featuring 106 billion total parameters with only 12 billion active per forward pass. This design incorporates 128 routed experts and a specialized shared expert layer, activating 9 experts per token to maintain representational capacity while significantly reducing computational overhead. The transformer block is further enhanced by a Multi-Token Prediction (MTP) layer, which allows the model to predict several future tokens simultaneously. This implementation facilitates speculative decoding, which increases inference throughput and provides a responsive experience for real-time interactive applications.

Technical innovations in GLM-4.5-Air include the adoption of Grouped-Query Attention (GQA) with 96 attention heads and 8 key-value groups, reducing memory bandwidth requirements during long-context processing. The model supports a 128,000-token context window using Rotary Positional Embeddings (RoPE) and features a hybrid reasoning system. This system allows for a deliberate thinking mode, which executes a latent chain-of-thought process for analytical problem-solving, and a standard mode for immediate output. Native integration for function calling, web browsing, and code execution ensures the model can interact with external environments with high reliability.